Generalization algorithm of multimodal pre-training model based on graph-text self-supervised training